Now instead of taking a function, they take two lists, one z, the other
bias(z). This is to prevent double interpolation (one by Python sampling a
non-analytic function, one by GSL), and it's probably more familiar to the
user.

I should make it possible to set either of those

I don't want to figure out how distutils packaging works just yet, for now it's
easiest to put everything in one enormous file and package that.

Forgot to set the power spectrum each time the contributions changed

This one was quite challenging to find; apparently, by calling
`set_power_spectrum_linear`, I modified k_min and k_max, but whenever I
re-computed it using CLASS, it was not being adjusted accordingly, so the
integrals would all give `nan`s, which of course propagated further and
caused the tests to fail. Hopefully, that's all I missed, seems to work now.

Something goes wrong with the `pk_type` setter, there's probably something
that gets freed improperly, or some value isn't changed when it should be.

Still need to put the in the actual Coffe class.

This is primarily useful for nonlinearities, where we can't just rescale the
power spectrum by the linear growth factor.

Since there's no other way to evaluate nonlinearities for now, there's no
point in having a toggle for the midpoint approximation.

Sadly, C doesn't have constructors so I usually forget to initialize
structures, which lead to segfaults since they are accidentally freed later.

This way I need to modify one function instead of hunting around for various
pieces in the parameters.

There are a couple of issues with the Zeldovich approximation:
- computing the integrals on-the-fly is ridiculously slow, even for density-density
- on the other hand, GSL doesn't have support for 3D interpolations
- the midpoint approximation has about the same accuracy (see arXiv:2011.06185),
  and only requires 2D interpolation
- lastly, the code wasn't using it anymore

Hence, I just removed the option altogether from everywhere.

Leaving in the paralellization causes the code to hand when running the OpenMP
loops, so that one has to be disabled then.  Running some tests, it seems the
parallel multidimensional sampling is actually slower than just using OpenMP,
so it'd be ideal to disable Cuba paralellization, which is what this commit
achieves by just setting the `CUBACORES` environmental variable to 0, which
works.
Strangely enough, if I directly call `cubacores` from the C code, and set
`CUBACORES` to 0, which should have the same effect, it crashes the Cython code
with a very strange error:

python3.7: ./src/common/Fork.c:100: cubafork: Assertion `socketpair(AF_LOCAL,
SOCK_STREAM, 0, fd) != -1 && (pid = fork()) != -1' failed.

I've tried experimenting with the Cython `nogil` option, but to no avail, so
this is more of a hack than anything (that works!).